This manuscript is for review purposes only.

1 2 3 4 5 6 7 8 9 10 11 12 THE USE OF QUADRATIC REGULARIZATION WITH A CUBIC DESCENT CONDITION FOR UNCONSTRAINED OPTIMIZATION E. G. BIRGIN AND J. M. MARTíNEZ Abstract. Cubic-regularization and trust-region methods with worst-case first-order complexity O(ε 3/2 ) and worst-case second-order complexity O(ε 3 ) have been developed in the last few years. In this paper it is proved that the same complexities are achieved by means of a quadraticregularization method with a cubic sufficient-descent condition instead of the more usual predictedreduction based descent. Asymptotic convergence and order of convergence results are also presented. Finally, some numerical experiments comparing the new algorithm with a well-established quadratic regularization method are shown. Key words. Unconstrained minimization, quadratic regularization, cubic descent, complexity. AMS subject classifications. 90C30, 65K05, 49M37, 90C60, 68Q25. 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 1. Introduction. Assume that f : R n R is possibly nonconvex and smooth for all x R n. We will consider the unconstrained minimization problem given by (1) Minimize f(x). In the last decade, many works have been devoted to analyze iterative algorithms for solving (1) from the point of view of their time complexity. See, for example, [2, 4, 5, 6, 8, 11, 14, 19, 21]. A review of complexity results for the convex case, in addition to novel techniques, can be found in [12]. Given arbitrary tolerances ε g > 0 and ε h > 0, the question is about the amount of iterations and functional and derivative evaluations that are necessary to achieve an approximate solution defined by f(x) ε g or by f(x) ε g plus λ 1 ( 2 f(x)) ε h, where λ 1 ( 2 f(x)) represents the left-most eigenvalue of 2 f(x). In general, gradient-based methods exhibit complexity O(ε 2 g ) [4], which means that there exists a constant c, that only depends on the characteristics of the problem, algorithmic parameters, and, of course, the initial approximation, such that the effort required to achieve f(x) ε g for a bounded-below objective function f is at most c/ε 2 g. This bound is sharp for all gradient-based methods [4]. Complexity results for modified Newton s methods are available in [14]. Surprisingly, Newton s method with the classical trust-region strategy does not exhibit better complexity than O(ε 2 g ) either [4]. The same example used in [4] to prove this fact can be applied to Newton s method with standard quadratic regularization. On the other hand, Newton s method employing cubic regularization [15] for obtaining sufficient descent at each iteration exhibits the better complexity O(ε 3/2 g ) (see [5, 6, 19, 21]). The best known practical algorithm for unconstrained optimization with worstcase evaluation complexity O(ε 3/2 g ) to achieve first-order stationarity and complexity O(ε 3/2 g + ε 3 h ) to achieve second-order stationarity, defined by Cartis, Gould, and Submitted to the editors October 27, 2016. Funding: Supported by FAPESP (grants 2013/03447-6, 2013/05475-7, 2013/07375-0, and 2014/18711-3) and CNPq (grants 309517/2014-1 and 303750/2014-6). Dept. of Computer Science, Institute of Mathematics and Statistics, University of São Paulo, Rua do Matão, 1010, Cidade Universitária, 05508-090, São Paulo, SP, Brazil. egbirgin@ime.usp.br Dept. of Applied Mathematics, Institute of Mathematics, Statistics, and Scientific Computing, State University of Campinas, 13083-859, Campinas, SP, Brazil. martinez@ime.unicamp.br 1

2 E. G. BIRGIN AND J. M. MARTíNEZ 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 Toint in [5] and [6], uses cubic regularization and a descent criterion based on the comparison of the actual reduction of the objective function and the reduction predicted by a quadratic model. A non-standard trust-region method with the same complexity properties due to Curtis, Robinson, and Samadi [8] employs a cubic descent criterion for accepting trial increments. In [2], the essential ideas of ARC [5, 6] were extended in order to introduce high-order methods in which a p-th Taylor approximation (p 2) plus a (p + 1)-th regularization term is minimized at each iteration. In these methods, O(ε (p+1)/p g ) evaluation complexity for first-order stationarity is obtained also using the actual-versus-predicted-reduction descent criterion. However, it is rather straightforward to show that this criterion can be replaced by a (p+1)-th descent criterion (i.e. f(x k+1 ) f(x k ) α x k+1 x k p+1 ) in order to obtain the same complexity results. Moreover, the (p + 1)-th descent criterion (cubic descent in the case p = 2) seems to be more naturally connected with the Taylor approximation properties that are used to prove complexity. Cubic descent was also used in [19] in a variable metric method that seeks to achieve good practical global convergence behavior. In the trust-region example exhibited in [4], the unitary Newtonian step is accepted at every iteration since it satisfies the adopted sufficient descent criterion. This criterion requires that the function descent (actual reduction) should be better than a fraction of the predicted descent provided by the quadratic model (predicted reduction). However, if, instead of this condition, one requires functional descent proportional to s 3, where s is the increment given by the model minimization, the given example does not stand anymore. This state of facts led us to the following theoretical question: Would it be possible to obtain worst-case evaluation complexities O(ε 3/2 g ) and O(ε 3/2 g + ε 3 h ) using cubic descent to accept trial increments but only quadratic regularization in the subproblems? In this paper, we provide an affirmative answer to this question by incorporating cubic descent into a quadratic regularization framework. Iterative regularization is a classical idea in unconstrained optimization originated in the seminal works of Levenberg [17] and Marquardt [18] for nonlinear least-squares. It relies upon the Levenberg-Marquardt path, which is the set of solutions of regularized subproblems varying the regularization parameter, both in the case of quadratic and cubic regularized subproblems. It is worth mentioning that this path is also the set of solutions of Euclidean trust-region subproblems for different trust-region radii. The explicit consideration of the so-called hard case (where the Hessian is not positive definite and the gradient is orthogonal to the eigenspace related to the left-most Hessian s eigenvalue) and the employment of spectral computations to handle it are in the core of every careful trust-region implementation [8, 20, 22, 23]. Our new method explicitly deals with the hard case and uses a regularization parameter with adequate safeguards in order to guarantee the classical complexity results of cubic regularization and related methods [8]. The new method has been implemented and compared against a well established quadratic regularization method for unconstrained optimization introduced in [16]. The rest of this paper is organized as follows. A model algorithm with cubic descent is described in section 2. An implementable version of the algorithm is introduced in section 3. Well-definiteness and complexity results are presented in section 4 and section 5, respectively. Local convergence results are given in section 6. Numerical experiments are presented in section 7; while final remarks are given in section 8. Notation. The symbol denotes the Euclidean norm of vectors and the subordinate matricial norm. We denote g(x) = f(x), H(x) = 2 f(x), and, some-

QUADRATIC REGULARIZATION WITH A CUBIC DESCENT CONDITION 3 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 times, g k = g(x k ) and H k = H(x k ). If a R, [a] + = max{a, 0}. If a 1,..., a n R, diag(a 1,..., a n ) denotes the n n diagonal matrix whose diagonal entries are a 1,..., a n. If A R n n, A denotes de Moore-Penrose pseudoinverse of A. The notation [x] j denotes the jth component of a vector x whenever the simpler notation x j might lead to confusion. 2. Model algorithm. The following algorithm establishes a general framework for minimization schemes that use cubic descent. At each iteration k, we compute an increment s k such that f(x k +s k ) f(x k ) α s k 3. In principle, this is not very useful because even s k = 0 satisfies this descent condition. However, in Theorem 2.1, we show that under the additional condition (3), the algorithm satisfies suitable stopping criteria. As a consequence, practical algorithms should aim to achieve (2) and (3) simultaneously. Algorithm 2.1. Let x 0 R n and α > 0 be given. Initialize k 0. Step 1. Compute s k such that (2) f(x k + s k ) f(x k ) α s k 3. Step 2. Define x k+1 = x k + s k, set k k + 1, and go to Step 1. The theorems below establish that, under suitable assumptions, every limit point of the sequence generated by Algorithm 2.1 is second-order stationary and provide an upper bound on the number of iterations that Algorithm 2.1 requires to achieve a target objective functional value or to find an approximate first- or second-order stationary point. Lemma 2.1. Assume that the objective function f is twice continuously differentiable and that there exist γ g > 0 and γ h > 0 such that, for all k N, the increment s k computed at Step 1 of Algorithm 2.1 satisfies (3) g k+1 s k and [ λ 1,k] + s k, γ g γ h where λ 1,k stands for the left-most eigenvalue of H k. Then, it follows that f(x k+1 ) f(x k ) max {( α γ 3/2 g ) ( ) } α g k+1 3/2, γh 3 [ λ 1,k ] 3 +. 114 115 Proof. The result follows trivially from (2), (3), and the fact that, at Step 2 of Algorithm 2.1, x k+1 is defined as x k+1 = x k + s k. 116 Theorem 2.1. Let f min R, ε g > 0, and ε h > 0 be given constants, assume 117 that the hypothesis of Lemma 2.1 hold, and let {x k } k=0 be the sequence generated by 118 Algorithm 2.1. Then, the cardinality of the set of indices 119 (4) K g = { k N f(x k ) > f min and g k+1 > ε g } 120 121 is, at most, (5) ( ) 1 f(x 0 ) f min ; α (ε g /γ g ) 3/2

4 E. G. BIRGIN AND J. M. MARTíNEZ 122 123 124 125 while the cardinality of the set of indices (6) K h = { k N f(x k } ) > f min and λ 1,k < ε h is, at most, (7) ( ) 1 f(x 0 ) f min α (ε h /γ h ) 3. 129 126 Proof. From Lemma 2.1, it follows that at every time an iterate x k is such that 127 g k+1 > ε g the value of f decreases at least α(ε g /γ g ) 3/2 ; while at every time an 128 iterate x k is such that λ 1,k < ε h the value of f decrease at least α(ε h /γ h ) 3. The thesis follows from the fact that, by (2), {f(x k )} k=0 is a non-increasing sequence. 130 Corollary 2.1. Let f min R, ε g > 0, and ε h > 0 be given constants and assume 131 that the hypothesis of Lemma 2.1 hold. Algorithm 2.1 requires O(εg 3/2 ) iterations to 132 compute x k such that 133 f(x k ) f min or g k+1 ε g ; 134 135 136 it requires O(ε 3 h ) iterations to compute x k such that and it requires O(ε 3/2 g f(x k ) f min or λ 1,k ε h ; + ε 3 h ) iterations to compute x k such that f(x k ) f min or ( g k+1 ε g and λ 1,k ε h ). 137 Corollary 2.2. Assume that the hypothesis of Lemma 2.1 hold and let {x k } 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 be the sequence generated by Algorithm 2.1. bounded below, we have that lim k g(xk ) = 0 and k=0 Then, if the objective function f is lim [ λ 1,k] + = 0. k Proof. Assume that lim k g(x k ) = 0. This means that there exists ε > 0 and K, an infinite subsequence of N, such that g k > ε for all k K. Since f is bounded below, this contradicts Theorem 2.1. The second part is analogous. Corollary 2.3. Assume that the hypothesis of Lemma 2.1 hold. Then, if the objective function f is bounded below, every limit point x of the sequence {x k } k=0 generated by Algorithm 2.1 is such that f(x ) = 0 and 2 f(x ) is positive semidefinite. Proof. This corollary follows from Corollary 2.2 by continuity of f and 2 f. 3. Implementable algorithm. Algorithm 2.1 presented in the previous section is a model algorithm in the sense that it does not prescribe a way to compute the step s k satisfying (2) and (3). This will be the subject of the present section. Algorithm 3.1 is almost identical to Algorithm 2.1 with the sole difference that it uses Algorithm 3.2 to compute s k. Lemma 4.1 shows that Algorithm 3.2 is well defined and Lemma 4.4 shows that the step s k computed by Algorithm 3.2 satisfies the hypothesis (3) of Lemma 2.1. In the following section, it will be shown that Algorithm 3.2 computes s k using O(1) evaluations of f (and a single evaluation of g and H at the current iterate x k ). This implies that the complexity results on the number of iterations of the model Algorithm 2.1 also apply to the number of iterations and evaluations of f and its first- and second-order derivatives performed by Algorithm 3.1 3.2.

QUADRATIC REGULARIZATION WITH A CUBIC DESCENT CONDITION 5 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 Algorithm 3.1. Let x 0 R n, α > 0, and M > 0 be given. Initialize k 0. Step 1. Use Algorithm 3.2 to compute s R n satisfying (8) f(x k + s) f(x k ) α s 3 and define s k = s. Step 2. Define x k+1 = x k + s k, set k k + 1, and go to Step 1. Algorithm 3.2 below describes the way in which the increment s k is computed. For that purpose, different trial increments are tried along the set of solutions (9) s(µ) := argmin g k, s + 1 2 st ( H k + [ λ 1,k ] + I ) s + µ 2 s 2, for different values of the regularizing parameter µ 0, where λ 1,k is the left-most eigenvalue of H k. Algorithm 3.2 proceeds by increasing the value of the regularization parameter µ 0 until the sufficient descent condition (8) is satisfied with s = s(µ). For each value of µ, we define ρ(µ) = ([ λ 1,k ] + + µ)/(3 s(µ) ). By Lemma 3.1 of [5] (see also [15, 21]), s(µ) is a global minimizer of g k, s + 1 2 st H k s + ρ(µ) s 3. The way in which µ is increased is determined by two necessities related to ρ(µ): the initial ρ(µ) at each iteration should not be excessively small and the final ρ(µ) should not be excessively big. Essentially, the technical manipulation of the quadratic regularization parameter µ in the algorithm is motivated by these two apparently conflicting objectives which are necessary to obtain the complexity results. Algorithm 3.2. Given x k, this algorithm computes a step s R n satisfying (8). Step 1. Let λ 1,k be the left-most eigenvalue of H k. Consider the linear system (10) [H k + ([ λ 1,k ] + + µ)i]s = g k. If (10) with µ = 0 is not compatible then set ρ k,0 = 0 and go to Step 5; else pursue to Step 2 below. Step 2. Compute the minimum norm solution ŝ k,0 to the linear system (10) with µ = 0 and set, if ŝ k,0 = 0 and [ λ 1,k ] + > 0, ρ k,0 = 0, if ŝ k,0 = 0 and [ λ 1,k ] + = 0, [ λ 1,k ] + / ( 3 ŝ k,0 ), if ŝ k,0 0. If ρ k,0 M then go to Step 4; else pursue to Step 3 below. Step 3. Let q 1,k with q 1,k = 1 be an eigenvector of H k associated with its left-most eigenvalue λ 1,k. Set l 3 1 and compute t l3 0 and ŝ k,l3 = ŝ k,0 + t l3 q 1,k such that (11) [ λ 1,k ] + / ( 3 ŝ k,l3 ) = M. If (8) holds with s = ŝ k,l3, return s = ŝ k,l3 ; else pursue to Step 3.1 below. Step 3.1. While ŝ k,l3 2 ŝ k,0, execute Steps 3.1.1 3.1.2 below: Step 3.1.1. Set l 3 l 3 + 1 and compute t l3 0 and ŝ k,l3 = ŝ k,0 + t l3 q 1,k such that (12) ŝ k,l3 = 1 2 ŝk,l3 1. Step 3.1.2. If (8) holds with s = ŝ k,l3 then return s = ŝ k,l3.

6 E. G. BIRGIN AND J. M. MARTíNEZ 199 200 201 Step 4. If (8) holds with s = ŝ k,0 then return s = ŝ k,0 ; else pursue to Step 5 below. Step 5. Set l 5 1 and ρ k,l5 = max{0.1, ρ k,0 } and compute µ k,l5 > 0 and s k,l5 solution to (10) with µ = µ k,l5 such that 202 (13) ρ k,l5 [ λ 1,k] + + µ k,l5 3 s k,l5 100ρ k,l5. 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 If (8) holds with s = s k,l5, return s = s k,l5 ; else pursue to Step 5.1 below. Step 5.1. While µ k,l5 < 0.1, execute Steps 5.1.1 5.1.3 below: Step 5.1.1. Set l 5 l 5 + 1 and ( ) [ λ1,k ] + + µ k,l5 1 (14) ρ k,l5 = 10 3 s k,l5 1. Step 5.1.2 Compute µ k,l5 > 0 and s k,l5 solution to (10) with µ = µ k,l5 such that (13) holds. Step 5.1.3 If (8) holds with s = s k,l5, return s = s k,l5. Step 6. Set l 6 1, µ k,l6 = 2 µ k,l5, and compute s k,l6 solution to (10) with µ = µ k,l6. Step 6.1. While (8) does not hold with s = s k,l6, execute Steps 6.1.1 6.1.2 below: Step 6.1.1. Set l 6 l 6 + 1 and µ k,l6 = 2 µ k,l6 1. Step 6.1.2. Compute s k,l6 solution to (10) with µ = µ k,l6. Step 6.2. Return s = s k,l6. The reader may have noticed that Algorithm 3.2 includes several constants in its definition. Those constants are arbitrary and all of them can be replaced by any number (sometimes larger or smaller than unity, depending on the case). The algorithm was presented in this way with the simple purpose of avoiding a large number of hard-to-recall letters and/or parameters. The way in which Algorithm 3.2 proceeds is directly related to the geometry of the set of solutions of (9), many times called Levenberg-Marquardt path. On the one hand, when µ, s(µ) tends to 0 describing a curve tangent to g k. On the other hand, the geometry of the Levenberg-Marquardt path when µ 0 depends on the positive definiteness of H k and the compatibility or not of the linear system (10) with µ = 0 as we now describe. If H k is positive definite then the Levenberg-Marquardt path is a bounded curve that joins s = 0 with the Newtonian step s = (H k ) 1 g k. In this case, we have that λ 1,k > 0, so [ λ 1,k ] + = 0. Then, the system (10) with µ = 0 is compatible and, by Step 2, ρ k,0 = 0. Since ρ k,0 M, the algorithm continues at Step 4 and the increment ŝ k,0 is accepted if the sufficient descent condition (8) holds with s = ŝ k,0 (this is always the case if ŝ k,0 = 0, that occurs if and only if g k = 0). However, if (8) does not hold, after a few initializations at Step 5, the algorithm computes at Step 5.1.2 a regularization parameter µ such that the corresponding ρ(µ) increases with respect to the previous one, but not very much. This corresponds to our purpose of maintaining the auxiliary quantity ρ(µ) within controlled bounds. If s(µ) does not satisfy (8) (checked at Step 5.1.3) and the regularization parameter µ is still small (checked at the loop condition of Step 5.1), we update (increase) the bounds on ρ(µ) at Step 5.1.1, and we repeat this process until the fulfillment of (8) or until µ is not small anymore. In that latter case, the process continues in Step 6 with regular increases of the regularization parameter µ which should lead to the final fulfillment of (8) at the loop condition of Step 6.1. It is easy to see that, when H k is positive semidefinite and the linear system H k s = g k is compatible, the algorithm proceeds as in the positive definite case described above.

QUADRATIC REGULARIZATION WITH A CUBIC DESCENT CONDITION 7 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 The case in which H k is not positive definite but the linear system (10) with µ = 0 is compatible is called the hard case in the trust-region literature [7]. In the hard case, the Levenberg-Marquardt path is constituted by two branches. The first branch, that corresponds to µ > 0, is a bounded curve that joins s = 0 with the minimumnorm solution of (10) with µ = 0. The second branch, that corresponds to µ = 0, is given by the infinitely many solutions to the system (10) with µ = 0. This set of infinitely many solutions form an affine subspace that contains [H k + [ λ 1,k ] + I] g k and is spanned by the eigenvectors of H k associated with λ 1,k. Usually, one restricts this affine subspace to the line [H k + [ λ 1,k ] + I] g k + tv with t R, where v is one of the eigenvectors associated with λ 1,k. The algorithm starts by computing the minimum norm solution of (10) with µ = 0, which corresponds to the intersection of the two branches of the Levenberg-Marquardt path. If taking the regularizing parameter µ = 0 we have that the associated ρ(µ) is not very big (ρ k,0 M at Step 2) then we proceed exactly as in the positive definite and compatible positive semidefinite cases, increasing µ and seeking an acceptable increment along the first branch of the Levenberg-Marquardt path. However, if ρ k,0 > M, we are in the case in which ρ(µ) could be very big. Then, the search starts at Step 3 by seeking an increment along the second branch of the Levenberg-Marquardt path. This happens when λ 1,k < 0 and ŝ k,0 = 0 (because g k = 0), since in that case, we set ρ k,0 = at Step 2. Note that, along this branch, the value of µ = 0 does not change and the reduction of ρ(µ) is achieved trivially by increasing the norm of s(µ). Starting with a sufficiently large s(µ), and by means of successive reductions of s(µ) at Step 3.1.1, we seek the fulfillment of (8). However, after a finite number of reductions of s(µ) this norm becomes smaller than a multiple of the norm of the minimum-norm solution (except in the case in which we have ŝ k,0 = 0). If this happens, we enter Step 4 and then initiate a search in the other branch in an analogous way as we do in the positive definite case. In this situation, we have the guarantee that ρ(µ) is suitable bounded in the intersection point because, otherwise, the sufficient descent condition (8) would have been satisfied. If H k is not positive definite and the system (10) with µ = 0 is not compatible then the Levenberg-Marquardt path is an unbounded curve that, as µ tends to 0, becomes tangent to an affine subspace generated by an eigenvector of H k associated with λ 1,k. In this case, the control goes to Step 5 and the algorithm proceeds as in the already described situation in which H k is positive definite but the Newtonian step does not satisfy the sufficient descent condition (8). 279 280 281 282 283 284 285 286 287 288 289 290 291 4. Well-definiteness results. In this section, we will show that Algorithm 3.2 is well-defined and that the computed increment s k that satisfies (8) also satisfies (3). We start by describing how Algorithm 3.2 could be implemented considering the spectral decomposition of H k. Of course, this is an arbitrary choice and other options are possible like, for example, computing the left-most eigenvalue of H k only, and possible its associated eigenvector, and then solving the linear systems by any factorization suitable for symmetric matrices. In any case, the description based on the spectral decomposition of H k introduces some useful notation for the rest of the section. Consider the spectral decomposition H k = Q k Λ k Q T k, where Q k = [q 1,k... q n,k ] is orthogonal and Λ k = diag(λ 1,k,..., λ n,k ) with λ 1,k λ n,k. Substituting H k by its spectral decomposition in (10), we obtain [Λ k + ([ λ 1,k ] + + µ)i]q T k s = QT k gk. Therefore, for µ = 0, the linear system (10) is compatible if and only if [Q T k gk ] j = 0 whenever λ j,k + [ λ 1,k ] + = 0. Assuming that the linear system (10) with µ = 0 is

8 E. G. BIRGIN AND J. M. MARTíNEZ 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 compatible, its minimum norm solution is given by ŝ k,0 = Q k y k, where { yj k [Q T = k g k ] j /(λ j,k + [ λ 1,k ] + ), j J, 0, j J, J = {j {1,..., n} λ j,k + [ λ 1,k ] + 0}, and J = {1,..., n} \ J. Moreover, note that ŝ k,0 ( = [Q T k g k ] j /(λ j,k + [ λ 1,k ] + ) ) 2. j J The norm of ŝ k,l3 = ŝ k,0 + t l3 q 1,k (for any l 3 1) computed at Step 3 is given by ŝ k,l3 = ŝ k,0 2 + t l3 2 ŝ k,0, q 1,k + t 2 l 3 = ŝ k,0 2 + t 2 l 3, where the last equality holds because ŝ k,0 is orthogonal to q 1,k by definition. Thus, given a desired norm c l3 for ŝ k,l3 (c l3 = [ λ 1,k ] + /(3M) when l 3 = 1 and c l3 = 1 2 ŝk,l3 1 when l 3 > 1), we have that t l3 = c 2 l 3 ŝ k,0 2. The following technical lemma establishes that Step 5 of Algorithm 3.2 can always be completed finding a regularization parameter µ and an increment s(µ) that satisfies (13). The assumption g k 0 in the lemma is perfectly reasonable because, as it will be shown later, it always holds at Step 5. Lemma 4.1. Suppose that g k 0. At Step 5 of Algorithm 3.2, for any l 5 1, there exists µ k,l5 > 0 and s k,l5 solution to (10) with µ = µ k,l5 satisfying (13). Proof. For any µ > 0, the matrix of the system (10) is positive definite and the solution s(µ) to (10) is such that ( [Q (15) s(µ) = T ) 2 k gk ] j. (λ j,k + [ λ 1,k ] + + µ) Moreover, clearly, {j [Q T k gk ] j 0} (16) lim s(µ) = 0. µ In order to analyze the case µ 0, the proof will be divided in two cases: (a) the linear system (10) with µ = 0 is compatible and (b) the linear system (10) with µ = 0 is not compatible. Consider first case (a). In this case, since [Q T k gk ] j = 0 whenever λ j,k +[ λ 1,k ] + = 0, (15) is equivalent to s(µ) = ( [Q T ) 2 k gk ] j. (λ j,k + [ λ 1,k ] + + µ) j J Therefore, (17) lim µ 0 s(µ) = ŝ k,0 > 0 because g k 0 implies ŝ k,0 0. Thus, by (16) and (17), we have that [ λ 1,k ] + + µ [ λ 1,k ] + + µ (18) lim = and lim = [ λ 1,k] + µ 3 s(µ) µ 0 3 s(µ) 3 ŝ k,0.

QUADRATIC REGULARIZATION WITH A CUBIC DESCENT CONDITION 9 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 Since, by definition, for any l 5 1, ρ k,l5 ρ k,0 = [ λ 1,k] + 3 ŝ k,0, the desired result follows by continuity from (18). Consider now case (b). In this case, there exists j such that λ j,k + [ λ 1,k ] + = 0 and [Q T k gk ] j 0. Therefore, from (15), we have that (19) lim µ 0 s(µ) =. Thus, by (16) and (19), we have that [ λ 1,k ] + + µ [ λ 1,k ] + + µ (20) lim = and lim = 0. µ 3 s(µ) µ 0 3 s(µ) Since, by definition, for any l 5 1, in this case we have ρ k,l5 ρ k,0 = 0.1, the desired result follows by continuity from (20). Below we state the main assumption that supports the complexity results. Essentially, we will assume that the objective function is twice continuously differentiable and that 2 f satisfies a Lipschitz condition on a suitable region that contains the iterates x k and the trial points x k + s trial. Of course, a sufficient condition for the fulfillment of this assumption is the Lipschitz-continuity of 2 f on R n, but in some cases this global assumption may be unnecessarily strong. Assumption A1. The function f is twice continuous differentiable for all x R n and there exists a constant L > 0 such that, for all x k computed by Algorithm 3.1 and every trial increment s trial computed at Steps 2, 3, 3.1.1, 5, 5.1.2, 6, or 6.1.2 of Algorithm 3.2, we have that and f(x k + s trial ) f(x k ) + (s trial ) T g k + 1 2 (strial ) T H k s trial + L s trial 3 g(x k + s trial ) g k H k s trial L s trial 2. In the following lemma we prove that any trial increment necessarily satisfies the sufficient descent condition (8) if the regularization parameter is large enough. Lemma 4.2. Suppose that Assumption A1 holds and µ 0. If 0 s trial R n computed at Steps 2, 3, 3.1.1, 5, 5.1.2, 6, or 6.1.2 of Algorithm 3.2, that by definition satisfies (21) [H k + ([ λ 1,k ] + + µ)]s trial = g k, 352 353 is such that (22) [ λ 1,k ] + + µ 3 s trial L + α 354 355 356 then (8) is satisfied with s = s trial. Proof. Let us define, for all s R n, q(s) = s T g k + 1 2 st H k s.

10 E. G. BIRGIN AND J. M. MARTíNEZ 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 Since H k + ([ λ 1,k ] + + µ)i is positive semidefinite for any µ 0, by (21), (23) s trial minimizes q(s) + 1 2 ([ λ 1,k] + + µ) s 2. Define (24) ρ = [ λ 1,k] + + µ 3 s trial. By Lemma 3.1 of [5], s trial is a minimizer of q(s) + ρ s 3. In particular, (25) q(s trial ) + ρ s trial 3 q(0) = 0. Now, by Assumption A1, we have that f(x k + s trial ) f(x k ) + (s trial ) T g k + 1 2 (strial ) T H k s trial + L s trial 3 = f(x k ) + q(s trial ) + ρ s trial 3 + (L ρ) s trial 3. Thus, by (22), (24), and (25), f(x k + s trial ) f(x k ) α s trial 3. This completes the proof. The lemma below shows that Algorithm 3.2 may return a null increment only at Step 4. Lemma 4.3. Suppose that A1 holds. Algorithm 3.2 returns a null increment s = 0 if and only if g k = 0 and λ 1,k 0. Moreover, an increment s = 0 may only be returned by Algorithm 3.2 at Step 4 (i.e. Steps 3, 3.1.2, 5, 5.1.3, and 6.2 always return non null increments). Proof. Assume that g k = 0 and λ 1,k 0. Then, we have that the minimum norm solution ŝ k,0 to the linear system (10) with µ = 0 computed at Step 2 is null and that ρ k,0 = 0 M. Therefore, the algorithm goes to Step 4 and returns s = ŝ k,0 = 0 since it satisfies (8). Assume now that Algorithm 3.2 returned an increment s = 0. Since every trial increment computed by the algorithm is a solution to the linear system (10) for some µ 0, we must have g k = 0. If λ 1,k 0, the first part of thesis holds and it remains to show that the null increment is returned at Step 4. Note that, since g k = 0 implies ŝ k,0 = 0 and λ 1,k 0 means [ λ 1,k ] + = 0, at Step 2 we have ρ k,0 = 0 M. Thus, the algorithm goes to Step 4 where the null increment is returned since it satisfies (8). We now show that assuming λ 1,k < 0 leaves to a contradiction. Since λ 1,k < 0 means [ λ 1,k ] + > 0 and g k = 0 implies ŝ k,0 = 0, by the way ρ k,0 is defined at Step 2, we have that ρ k,0 = M. In this case the algorithm goes to Step 3. On the one hand, note that ŝ k,0 = 0 implies that the algorithm never leaves the loop in Step 3.1 becasue its condition reduces to ŝ k,l3 0. On the other hand, note that, by halving the norm of the trial increments ŝ k,l3, since µ = 0 is fixed, in a finite number of trials, (22) holds and, by Lemma 4.2, the algorithm returns s = ŝ k,l3 0 for some l 3 1, contradicting the fact that the algorithm returned a null increment. We finish this section proving that the increment s k computed at Algorithm 3.2, that satisfies (8) and defines x k+1 in Algorithm 3.1, is such that it also satisfies (3). Note that this result assumes the existence of s k by hypothesis. Up to the present moment we proved that Algorithm 3.2 is well defined. The existence of s k for all k will be proved in the following section when proving that Algorithm 3.2 always computes s k performing a finite number of operations.

QUADRATIC REGULARIZATION WITH A CUBIC DESCENT CONDITION 11 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 Lemma 4.4. Suppose that Assumption A1 holds. Then, there exist γ g > 0 and γ h > 0 such that, for all k N, the increment s k computed by Algorithm 3.2 and the new iterate x k+1 = x k + s k computed at Step 2 of Algorithm 3.1 satisfy g k+1 s k and [ λ 1,k] + s k. γ g γ h Moreover, (26) γ g max {3M + L, 3000(L + α) + L, 30 + L} and (27) γ h max {3M, 3000(L + α), 30}. Proof. If s k = 0 then, by Lemma 4.3, we have that g k = 0 and λ 1,k 0 and, therefore, the thesis follows trivially. We now assume s k 0. Since s k is a solution to (10) for some µ 0, we have that H k s k + g k + ([ λ 1,k ] + + µ)s k = 0. Therefore, ( ) H k s k + g k [ λ1,k ] + + µ + s k s k s k = 0. Then ( ) H k s k + g k [ λ1,k ] + + µ = s k s k 2. But, by Assumption A1 and the triangle inequality, g k+1 g k + H k s k g k+1 g k H k s k L s k 2. Therefore, ( (28) g k+1 [ λ1,k ] + + µ s k ) + L s k 2. We now analyze in separate the cases in which s k 0 is returned by Algorithm 3.2 at Steps 3, 3.1.2, 4, 5, 5.1.3, and 6.2 Case s k = ŝ k,l3 with l 3 = 1 was returned at Step 3: In this case, s k,l3 is a solution to (10) with µ = 0 and, by (11), it satisfies (29) [ λ 1,k ] + / s k,l3 = 3M. Case s k = ŝ k,l3 with l 3 > 1 was returned at Step 3.1.2: This means that there exists ŝ k,l3 1 0 that is a solution to (10) with µ = 0 and for which (8) with s = ŝ k,l3 1 did not hold. Therefore, by Lemma 4.2, we have that [ λ 1,k ] + / ( 3 ŝ k,l3 1 ) < L + α. Thus, by (12), we have that (30) [ λ 1,k ] + / ŝ k,l3 < 6(L + α). Case s k = ŝ k,0 was returned at Step 4: In this case, we have that (31) [ λ 1,k ] + / ( 3 ŝ k,0 ) M or that there exists ŝ k,l3 0 with l 3 1 such that (32) ŝ k,l3 < 2 ŝ k,0,

12 E. G. BIRGIN AND J. M. MARTíNEZ 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 ŝ k,l3 is a solution to (10) with µ = 0, and (8) did not hold with s = ŝ k,l3. Therefore, by Lemma 4.2, we have that (33) [ λ 1,k ] + / ( 3 ŝ k,l3 ) < L + α and, by (32) and (33), (34) [ λ 1,k ] + / ŝ k,0 < 6(L + α). Thus, by (31) and (34), (35) [ λ 1,k ] + / s k,0 max{3m, 6(L + α)}. Case s k = s k,l5 with l 5 = 1 was returned at Step 5: In this case there are two possibilities: the linear system (10) with µ = 0 is compatible or not. In the first case, ŝ k,0 was computed, ρ k,0 = [ λ 1,k ] + / ( 3 ŝ k,0 ), and, since (8) with s = ŝ k,0 did not hold, by Lemma 4.2, ρ k,0 < L + α. In the second case, we simple have that ρ k,0 = 0. Thus, by (13) and by the fact that, by definition, ρ k,1 = max{0.1, ρ k,0 }, in the first case, we have (36) [ λ 1,k ] + + µ k,l5 3 s k,l5 and, in the second case, we have (37) [ λ 1,k ] + + µ k,l5 3 s k,l5 Therefore, µ k,l5 0, (36), and (37) imply that 100ρ k,l5 = 100 max{0.1, ρ k,0 } max{10, 100(L + α)} 100ρ k,l5 = 100 max{0.1, 0} = 10. 447 (38) [ λ 1,k ] + s k,l5 [ λ 1,k] + + µ k,l5 s k,l5 max{30, 300(L + α)}. 448 449 450 Case s k = s k,l5 with l 5 > 1 was returned at Step 5.1.3: This means that there exists µ k,l5 1 > 0 and s k,l5 1 solution to (10) with µ = µ k,l5 1 for which (8) did not hold. Thus, by Lemma 4.2, 451 [ λ 1,k ] + + µ k,l5 1 3 s k,l5 1 < L + α. 452 453 454 Moreover, by (13) and (14), Thus, [ λ 1,k ] + + µ k,l5 3 s k,l5 100ρ k,l5 = 1000 ( ) [ λ1,k ] + + µ k,l5 1 3 s k,l5 1. 455 (39) [ λ 1,k ] + s k,l5 [ λ 1,k] + + µ k,l5 s k,l5 3000(L + α). 456 457 458 Case s k = s k,l6 was returned at Step 6.2: If l 6 = 1 then µ k,l6 = 2 µ k,l5 for some l 5 1 and the solution s k,l5 to (10) with µ = µ k,l5 is such that (8) with s = s k,l5 does not hold. Thus, by Lemma 4.2, 459 [ λ 1,k ] + + µ k,l5 3 s k,l5 < L + α.

QUADRATIC REGULARIZATION WITH A CUBIC DESCENT CONDITION 13 460 461 462 463 464 465 On the other hand, and since µ k,l6 = 2 µ k,l5, we have that (40) ( ) s k,l [Q 6 = T 2 ( k gk ] j j J = λ j,k +[ λ 1,k ] + + µ k,l6 j J ( ) [Q = T 2 ( k gk ] j j J 2( 1 2 (λ j,k+[ λ 1,k ] + )+ µ k,l5 ) j J ( ) = 1 [Q T 2 k gk ] j 2 j J = 1 λ j,k +[ λ 1,k ] + + µ k,l5 2 sk,l 5 > 0. ) [Q T 2 k gk ] j λ j,k +[ λ 1,k ] + +2 µ k,l5 [Q T k gk ] j 2(λ j,k +[ λ 1,k ] + + µ k,l5 ) Therefore, (41) [ λ 1,k ] + s k,l6 [ λ 1,k] + + µ k,l6 = [ λ 1,k] + + 2 µ k,l5 = 2 ( 1 2 [ λ ) 1,k] + + µ k,l5 s k,l6 s k,l6 s k,l6 2 ([ λ 1,k ] + + µ k,l5 ) 2 ([ λ ( ) 1,k] + + µ k,l5 ) [ λ1,k ] + + µ k,l5 s k,l6 1 = 4 < 12(L + α). 2 sk,l5 s k,l5 If l 6 > 1 then µ k,l6 = 2 µ k,l6 1 and the solution s k,l6 1 to (10) with µ = µ k,l6 1 is such that (8) with s = s k,l6 1 does not hold. Thus, by Lemma 4.2, ) 2 466 (42) [ λ 1,k ] + + µ k,l6 1 3 s k,l6 1 < L + α. 467 468 469 Moreover, µ k,l6 = 2 µ k,l6 1 implies, as shown above, that (43) s k,l6 1 2 sk,l6 1. Therefore, by (42) and (43), and since µ k,l6 0, we have that 470 (44) [ λ 1,k ] + s k,l6 [ λ 1,k] + + µ k,l6 s k,l6 < 12(L + α). 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 The desired result (27) follows from (29), (30), (35), (38), (39), (41), and (44); while (26) follows from the same set of inequalities plus (28). 5. Complexity results. In this section, complexity results on Algorithm 3.2 are presented. In particular, we show that the number of functional evaluations required to compute the increment s k using Algorithm 3.2 is O(1), i.e. it does not depend on ε g nor ε h. The section finishes establishing the complexity of Algorithm 3.1 3.2 in terms of the number of functional (and derivatives) evaluations. The sufficient condition (8) is tested at Steps 3, 3.1.2, 4, 5, 5.1.3, and 6.1. These are the only steps of Algorithm 3.2 in which the objective function is evaluated. Condition (8) is tested only once per iteration at Steps 3, 4, and 5. Therefore, in order to assess the worstcase evaluation complexity of Algorithm 3.2, we must obtain a bound for the number of executions of the remaining mentioned steps, namely, Steps 3.1.2, 5.1.3, and 6.1. Step 3.1 of Algorithm 3.2 describes the loop that corresponds to the hard case, in which we seek an increment along an appropriate eigenvector of H k. For each trial increment, f is evaluated and the condition (8) is tested (at Step 3.1.2). Therefore, it is necessary to establish a bound on the number of executions of Step 3.1.2. This is done in Lemma 5.1. Lemma 5.1. Suppose that Assumption A1 holds. If Step 3.1.2 of Algorithm 3.2 is executed, it is executed at most log 2 ((L + α)/m) + 1 times.

14 E. G. BIRGIN AND J. M. MARTíNEZ 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 Proof. By (11) when l 3 = 1 and by (12) when l 3 > 1, ŝ k,l3 0 for all l 3 1 and or, equivalently, ŝ k,l3 = { [ λ1,k ] + /(3M), l 3 = 1, ŝ k,l3 1 /2, l 3 > 1, (45) 2 l3 1 M = [ λ 1,k ] + / ( 3 ŝ k,l3 ). Thus, by Lemma 4.2, if (8) does not hold with s = ŝ k,l3 we must have 2 l3 1 M < L+α, i.e. l 3 log 2 ((L + α)/m) + 1 as we wanted to prove. Step 5.1 of Algorithm 3.2 describes a loop where one tries to find an initial sufficiently big regularization parameter. Each time the regularization parameter is increased one tests the condition (8) (at Step 5.1.3). Therefore, it is necessary to establish a bound on the number of evaluations that may be performed at Step 5.1.3. This is done in Lemma 5.2. Lemma 5.2. Suppose that Assumption A1 holds. If Step 5.1.3 of Algorithm 3.2 is executed, it is executed at most log 10 (L + α) + 2 times. Proof. For all l 5 1, when (8) is tested at Step 5.1.3 with s = s k,l5, s k,l5 is a solution to (10) with µ = µ k,l5 > 0 and satisfies (13). Therefore, by Lemma 4.3, s k,l5 0 and, thus, by Lemma 4.2, if (8) does not hold with s = s k,l5 we must have (46) ρ k,l5 < L + α. 507 508 509 On the other hand, since, by definition, ρ k,1 0.1 and, by (13) and (14), ρ k,l5 10ρ k,l5 1 for all l 5 2, we have that (47) ρ k,l5 10 l5 2 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 for all l 5 1. By (46) and (47), if (8) does not hold with s = s k,l5 we must have 10 l5 2 < L + α, i.e. l 5 log 10 (L + α) + 2 as we wanted to prove. Finally, at Step 6.1 we increase the regularization parameter by means of a doubling process ( µ k,l6 = 2 µ k,l6 1). This process guarantees, by Lemma 4.3 and Lemma 4.2, that the sufficient condition will eventually hold. In Lemma 5.3, we prove that the number of doubling steps is also bounded by a quantity that only depends on characteristics of the problem and algorithmic parameters. For proving this lemma, we need to assume boundedness of H k at the iterates generated by the algorithm. Note that, since f(x k+1 ) f(x k ) for all k, a sufficient condition for Assumption A2 is the boundedness of H(x) on the level set defined by f(x 0 ). Assumption A2. There exists a constant h max 0 such that, for all iterates x k computed by Algorithm 3.1, we have that H k h max. Lemma 5.3. Suppose that Assumption A1 and Assumption A2 hold. If Step 6.1.2 of Algorithm 3.2 is executed, it is executed at most [ ( )] 1 ( ) 0.2 L + α log 1 + log + 1 h max + 0.2 0.1 times.

QUADRATIC REGULARIZATION WITH A CUBIC DESCENT CONDITION 15 526 527 528 529 530 Proof. For all l 6 1, Lemma 4.3 implies that s k,l6 0 and straightforward calculations show that ( s k,l6 = [Q T k g k ] j /(λ j,k + [ λ 1,k ] + + µ k,l6 ) ) 2. j J Moreover, it is easy to see that s k,l6 decreases when µ k,l6 increases. Therefore, since, by definition, µ k,l6+1 = 2 µ k,l6, for all l 6 1, we have that 531 (48) s k,l6 s k,l6+1 1. 532 533 534 535 536 Thus, for all l 6 1, (49) ( [ λ1,k ] ++ µ k,l6 +1 3 s k,l 6 +1 [ λ 1,k ] ++ µ k,l6 +1 [ λ 1,k ] ++ µ k,l6 ) ( ) [ λ1,k ] / ++ µ k,l6 = 3 s k,l 6 = [ λ 1,k] ++2 µ k,l6 [ λ 1,k ] ++ µ k,l6 = 1 + ( [ λ1,k ] ++ µ k,l6 +1 [ λ 1,k ] ++ µ k,l6 µ k,l6 [ λ 1,k ] ++ µ k,l6 ) ( s k,l 6 ) s k,l 6 +1 ( ) 1 + 0.2 h max+0.2 > 1, where the first inequality follows from (48) and the second inequality follows from the fact that, by the definition of the algorithm, µ k,l6 0.2 and by Assumption A2. From (49) and the fact that, by the definition of the algorithm, l 6 = 1 implies 537 [ λ 1,k ] + + µ k,l6 3 s k,l6 0.1, 538 539 it follows that (50) [ λ 1,k ] + + µ k,l6 3 s k,l6 ( ) l6 1 0.2 0.1 1 + h max + 0.2 540 541 542 543 for all l 6 1. For all l 6 1, when (8) is tested at Step 6.1.2 with s = s k,l6, s k,l6 satisfies (10) with µ = µ k,l6 > 0. Therefore, by Lemma 4.2, if (8) does not hold with s = s k,l6 we must have, by (50), ( ) l6 1 0.2 0.1 1 + < L + α. h max + 0.2 544 545 546 This implies the desired result. We finish this section summarizing the complexity and asymptotic results on Algorithm 3.1 3.2. 547 Theorem 5.1. Let f min R, ε g > 0, and ε h > 0 be given constants, suppose that 548 Assumption A1 and Assumption A2 hold, and let {x k } k=0 be the sequence generated 549 by Algorithm 3.1 3.2. Then, the cardinality of the set of indices (51) K g = { k N f(x k ) > f min and g k+1 } 550 > ε g 551 552 is, at most, (52) ( ) 1 f(x 0 ) f min ; α (ε g /γ g ) 3/2

16 E. G. BIRGIN AND J. M. MARTíNEZ 553 554 while the cardinality of the set of indices (53) K h = { k N f(x k ) > f min and λ 1,k < ε h } 555 556 is, at most, (54) ( ) 1 f(x 0 ) f min α (ε h /γ h ) 3, 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 where constants γ g and γ h are as in the thesis of Lemma 4.4 (i.e. they satisfy (26) and (27), respectively). Proof. Assumption A1 and Assumption A2 imply, by Lemma 4.4, the the hypothesis of Lemma 2.1 hold. Therefore, since Algorithm 3.1 is a particular case of Algorithm 2.1, the thesis follows from Theorem 2.1. Corollary 2.1, Corollary 2.2, and Corollary 2.3 also hold for Algorithm 3.1 3.2 under the hypothesis of Theorem 5.1, the most significant result being the complexity rates that possess the same dependencies on ɛ g and ɛ h whether we consider iteration or evaluation complexity. Note that the number of iterations is a direct consequence of Theorem 5.1. On the other hand, Lemma 5.1, Lemma 5.2, and Lemma 5.3 show that, every time Algorithm 3.2 is used by Algorithm 3.1 to compute an increment s k, it performs O(1) evaluations of the objective function f; while, by definition, it performs a single evaluation of g and H. Thus, the evaluation complexity of Algorithm 3.1 3.2 coincides with its iteration complexity. 6. Local convergence. Note that if H k is positive definite then the minimum norm solution ŝ k,0 to the linear system (10) with µ = 0 computed at Step 2 of Algorithm 3.2 is given by ŝ k,0 = (H k ) 1 g k, i.e. ŝ k,0 is the Newton direction. Moreover, since, independently of having ŝ k,0 = 0 or ŝ k,0 0, λ 1,k > 0 implies that ρ k,0 = 0 M, in this case (H k positive definite) the algorithm goes directly to Step 4 and checks whether the Newton direction satisfies the sufficient cubic decrease condition (8). The lemma below shows that, if (55) holds then the Newton direction satisfies (8). (If λ 1,k > 0 and g k = 0 and, in consequence, s k,0 = 0, it is trivial to see that the (null) Newton direction satisfies (8) and there is nothing to be proved. Anyway, the lemma below covers this case as well.) Lemma 6.1. Suppose that Assumption A1 holds. If H k is positive definite and (55) g k 1 2(L + α) λ2 1,k then we have that the trial increment ŝ k,0 computed at Step 2 of Algorithm 3.2 is such that (8) holds with s = ŝ k,0. Proof. By Assumption A1, f(x k + ŝ k,0 ) f(x k ) + (ŝ k,0 ) T g k + 1 2 (ŝk,0 ) T H k ŝ k,0 + L ŝ k,0 3. Then, since ŝ k,0 = (H k ) 1 g k, f(x k + ŝ k,0 ) f(x k ) 1 2 (ŝk,0 ) T H k ŝ k,0 + L ŝ k,0 3.

QUADRATIC REGULARIZATION WITH A CUBIC DESCENT CONDITION 17 589 590 591 592 593 594 595 Therefore, (56) f(x k + ŝ k,0 ) f(x k ) 1 2 λ 1,k ŝ k,0 2 + L ŝ k,0 3. On the other hand, since ŝ k,0 = (H k ) 1 g k, we have that (57) ŝ k,0 = (H k ) 1 g k (H k ) 1 g k = 1 λ 1,k g k. Then, by (55), ŝ k,0 λ 1,k /(2(L + α)) or, equivalently, λ 1,k /2 + L ŝ k,0 α ŝ k,0. Therefore, multiplying by ŝ k,0 2 and adding f(x k ), we have that f(x k ) 1 2 λ 1,k ŝ k,0 2 + L ŝ k,0 3 f(x k ) α ŝ k,0 3. 596 and the thesis follows from (56). 597 In the next theorem, we use the classical local convergence result of Newton s 598 method plus continuity arguments (that imply that the hypothesis (55) always hold 599 in a neighborhood of a local minimizer with positive definite Hessian) to prove the 600 quadratic local convergence of Algorithm 3.1 3.2. 601 Assumption A3. Let x be a local minimizer of f. We say that this assumption 602 holds if H(x ) is positive definite with H(x ) 1 β and, in addition, there exist 603 r > 0 and γ > 0 such that H(x) H(x ) γ x x whenever x x r. 604 Theorem 6.1. Let x be a local minimizer of f at which Assumption A3 holds 605 1 and suppose that Assumption A1 also holds. Define δ 1 = min{r, 2βγ }. Then, there 606 exists δ (0, δ 1 ] such that 607 (58) H(x) 1 2β whenever x x δ 608 and such that, if x 0 x δ, the sequence {x k } k=0 generated by Algorithm 3.1 3.2 609 satisfies [ ] (59) g(x k 1 610 ) /(2β) 2, 2(L + α) 611 612 613 614 615 616 617 618 619 620 621 (60) x k+1 x 1 2 xk x, and x k+1 x βγ x k x 2 for all k = 0, 1, 2,.... Proof. By the classical Newton convergence theory (see, for example, [9, Th.5.2.1, p.90]), whenever x 0 x δ 1 the sequence generated by x k+1 = x k (H k ) 1 g k is well defined and satisfies (60) for all k 0. By continuity of g(x), since g(x ) = 0, there exists δ 2 (0, δ 1 ] such that whenever x k x δ 2 one has that (59) holds; while, by continuity of H(x), there exists δ (0, δ 2 ] such that whenever x x δ one has that (58) holds. On the other hand, by (59), if x k x δ, we have that [ ] g(x k 1 ) / (H k ) 1 2 2(L + α)

18 E. G. BIRGIN AND J. M. MARTíNEZ 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 and, since (H k ) 1 = 1/λ 1,k, g(x k ) 1 2(L + α) λ2 1,k. Thus, by Lemma 6.1 and the definition of Algorithm 3.2, we have that x k+1 is, in fact, defined by x k+1 = x k (H k ) 1 g k and, therefore, the thesis follows by an inductive argument. Theorem 6.2. Let x be a local minimizer of f at which Assumption A3 holds. Suppose also that Assumption A1 holds and, in addition, that x is a limit point of the sequence {x k } k=0 generated by Algorithm 3.1 3.2. Then, the whole sequence {xk } k=0 converges quadratically to x. Proof. Since x is a limit point, there exists k 0 such that x k0 x δ. Thus, the convergence of {x k } follows from Theorem 6.1 replacing x 0 with x k0. The following is a global non-flatness assumption that will allow us to prove a complexity result that takes advantage of local quadratic convergence. Assumption A4. Let δ > 0 be as in the thesis of Theorem 6.1. There exists κ > 0 such that, for all x k generated by Algorithm 3.1 3.2, if x k x > δ then g(x k ) > κ. Note that Assumption A4 holds under the uniform non-singularity assumption that says that for all k N and x [x k, x k+1 ], H(x) is nonsingular and H(x) 1 1/η. In fact, by the Mean Value Theorem, the uniform non-singularity assumption implies that, for all x k generated by Algorithm 3.1 3.2, g(x k ) η x k x. Theorem 6.3. Let f be bounded below and let x be a local minimizer of f at which Assumption A3 holds. Suppose also that Assumption A1, Assumption A2 and Assumption A4 hold, and, in addition, that x is a limit point of the sequence {x k } k=0 generated by Algorithm 3.1 3.2. Then, after a number of iterations k 0 = O(κ 3/2 ), where κ is as in Assumption A4 and it only depends on characteristics of the problem and algorithmic parameters, we obtain that x k x δ for all k k 0, where δ is as in the thesis of Theorem 6.1. Proof. By construction (see Theorem 6.1), δ only depends on characteristics of the problem. By Assumption A4, g(x k ) > κ for all k such that x k x > δ. Then, by Assumption A1, Assumption A2, and Theorem 5.1, after k 0 = O(κ 3/2 ) iterations, we obtain that g(x k0 ) κ, i.e. x k0 x δ. This implies, by Theorem 6.1, that x k x δ for all k k 0, as we wanted to prove. Theorem 6.4. Let f be bounded below and let x be a local minimizer of f at which Assumption A3 holds. Suppose also that Assumption A1, Assumption A2, and Assumption A4 hold, and, in addition, that x is a limit point of the sequence {x k } k=0 generated by Algorithm 3.1 3.2. Let ε g > 0 be a given constant. Then, in at most ˆk = O(log 2 ( log 2 (ε g ))) iterations we have that g(x k ) ε g for all k ˆk. Proof. By the Mean Value Theorem of Integral Calculus, we have that, for any k 0, [ 1 (61) g(x k+1 ) = 0 ] H(ξ k+1 (t))dt (x k+1 x ), where ξ k+1 (t) = x + t(x k+1 x ). By the triangle inequality, Theorem 6.1, and Theorem 6.3, since x k+1 x δ for

QUADRATIC REGULARIZATION WITH A CUBIC DESCENT CONDITION 19 663 664 665 666 667 668 669 670 671 672 673 674 all k k 0 implies ξ(t) x δ for all k k 0 and t [0, 1], we have that (62) H(ξ k+1 (t)) H(x ) H(ξ k+1 (t)) H(x ) γ ξ k+1 (t) x γδ for all k k 0 and t [0.1]. Therefore, by (61) and (62), [ 1 (63) g(x k+1 ) = 0 ] H(ξ k+1 (t))dt (x k+1 x ) ( H(x ) + γδ) x k+1 x for all k k 0. On the other hand, by the Mean Value Theorem of Integral Calculus, we have that, for any k 0, [ 1 1 x k x = H(ξ k (t))dt] g(x k ) where ξ k (t) = x + t(x k x ) 0 and, thus, by Theorem 6.1 and Theorem 6.3, since x k x δ implies ξ k (t) x δ for all k k 0 and t [0, 1], we have that (64) x k x 2β g(x k ) for all k k 0. Now, by (63), (64), Theorem 6.1, and Theorem 6.3, 675 (65) g(x k+1 ) ( H(x ) + γδ) x k+1 x βγ( H(x ) + γδ) x k x 2 4β 3 γ( H(x ) + γδ) g k 2 676 677 678 679 680 681 for all k k 0. Up to this point, we have that g k0 κ with k 0 = O(κ 3/2 ) and that, for all l 0, g(x k0+1+l ) c quad g k0+l 2, where κ and c quad = 4β 3 γ( H(x ) + γδ) depend only on characteristics of the problem and algorithmic parameters. This means that (66) g(x k0+1+l ) c l+1 quad g(xk0 ) 2l+1 c l+1 quad κ2l+1 for all l 0. 682 We now consider, with the simple purpose of simplifying the presentation, k 1 k 0, 683 k 1 = O(c 3/2 quad ), whose existence is granted by Assumption A1, Assumption A2, and 684 Theorem 5.1, such that gk 1 2 c 1 quad for all k k 1. Thus, (66) can be re-stated as 685 686 687 688 689 690 691 692 693 694 (67) g(x k1+1+l ) c l+1 quad g(xk1 ) 2l+1 cl+1 quad c 2l+1 quad ( ) 2 l+1 1 2 2l+1 for all l 0. 2 Thus, since 2 2l+1 ε g if and only if l log 2 ( log 2 (ε g ))+1, we have that g k ε g for all k k 1 + log 2 ( log 2 (ε g )) + 1. This implies the desired result recalling that k 1 does not depend on ε g. 7. Numerical experiments. We implemented Algorithm 3.1 3.2 in Fortran 90. At each iteration k, the spectral decomposition of matrix H k is computed by the Lapack [1] subroutine dsyev. At Step 5 and 5.1.2 of Algorithm 3.2, µ k,l5 > 0 and s k,l5 solution to (10) with µ = µ k,l5 such that (13) holds are computed using bisection. In the numerical experiments, we arbitrarily considered α = 10 8 and M = 10 3. It should be noted that these two parameters, as well as the other constants that